Make the evaluator's review weights configurable.

[match/match.git] / program / PMDefaults.hs

module PMDefaults where
import PMInstance
import PMConfig

import qualified NaiveMinCostFlow
import qualified CS2MinCostFlow

-- A default set of configuration values; see module PMConfig.
pmDefaults = PMConfig {

-- === Choose a min-cost flow implementation (timings on mattlaptop2) ===

-- A naive implementation that is slow for all but the smallest instances
-- (30s on a 20x50 example).
minCostFlow = NaiveMinCostFlow.minCostFlow,

-- Uses CS2 (http://www.igsystems.com/cs2/), which requires a license for
-- non-research use but is faster (<1s on a 20x50 example, 64s on a 60x500
-- example).  Configure the path to cs2.exe in CS2MinCostFlow.hs.  Remember to
-- compile CS2 with -DPRINT_ANS, or this won't work!
--minCostFlow = CS2MinCostFlow.minCostFlow,

-- The number of reviews each proposal should get.
reviewsEachProposal = 3,

-- === Interpretation of the preference values ===

prefIsExpert = \p -> p <= 10,
prefIsKnowledgeable = \p -> p <= 20,

prefIsBoring = \p -> p > 15,
prefIsVeryBoring = \p -> p > 25,

prefIsConflict = \p -> p >= 40,

-- === Tuning parameters for the matcher ===

-- The number of reviews by which a reviewer's load may exceed his/her target
-- load of (relativeLoad * ceiling(numProps * reviewsEachProposal /
-- totalRelativeLoad)).  For now this is an additive constant; perhaps it should
-- be proportional to the target load.
loadTolerance = 1,

-- Cost to overload by one review.
-- tx = 0 at target load, 1 at end of tolerance.
marginalLoadCost = \tx -> 1000 + tx*1000,

-- Cost to review a boring (or very boring) proposal.
-- lx = 0 at no load, 1 at target load.
marginalBoringCost = \lx -> 1000 + lx*1000,
-- Additional cost to review a very boring proposal.
marginalVeryBoringCost = \lx -> 1000 + lx*1000,

-- Cost to make a review.  Used by the evaluator too.
-- I'm using quadratic cost functions as a first attempt.
assignmentCost = \pref -> (numAsWt 10 + pref) ^ 2,

-- Bonus for a first knowledgeable or expert review.
knowledgeableBonus = 1000,

-- Bonus for an additional expert review.
expertBonus = 1000,

-- === Parameters for the random-instance generator ===

-- Number of topics.
numTopics = 20,

-- Exponent of the Zipf distribution used to choose topics for each proposal.
topicZipfExponent = -0.5,

-- === Parameters for the matching evaluator ===

-- The weights given to the best, ..., worst review of a proposal in evaluating
-- its "unhappiness".  Default is [reviewsEachProposal, ..., 1] since it's most
-- important for the best review to be good.
reviewEvalWeights = let rep = reviewsEachProposal pmDefaults in
	map numAsWt [rep, rep-1 .. 1]

}